When companies build controls for AI agents, do they focus more on giving orders or on watching what the agent actually does?

This explores which parts of human oversight the people building agents put first: upfront permissions, autonomy dials, stop buttons, live monitoring, or after-the-fact records. The corpus gives a fairly clear answer. Builders favor control and intervention over observability. One design pattern library lists permissions, autonomy control, and pause/stop as three separate patterns, each placed at a different stage of the interaction, and this reflects a design community that puts levers ahead of windows into the agent's behavior Should agent control be one setting or several separate controls?. One caveat: the collection has no large survey of builders ranking these features, so this reading comes from design catalogs and user studies rather than from builders reporting their own priorities.

The research suggests the emphasis on upfront control is mostly well placed. In a 48-person study of agents that operate a computer, having users review the agent's plan before it ran cut problematic actions by 76%. Once a risky action was actually underway, though, users blocked only 21% of them, whatever oversight setup they had Does oversight strategy prevent risky agent actions or just catch them?. A stop button is technically present, but people rarely press it in time. Part of the reason is what overseers pay attention to. They check whether the agent followed its own plan, not whether the action itself is safe Do overseers focus on agent execution or action safety?. So a well-built pause control can still fail, because the person watching is asking the wrong question.

Developers working with coding agents show the same pattern in daily practice. Their oversight falls into four kinds of work: setting things up in advance, planning together with the agent, monitoring in real time, and reviewing afterward. Much of the effort goes in before the agent starts, configuring it and seeding solutions How do developers actually oversee AI coding agents?. When they do review results, they rely on shortcuts. Passing tests stand in for correctness, a sensible plan stands in for working code, and skimming replaces line-by-line reading Do developers actually review agent work exhaustively?. Builders seem to have worked out from experience that human attention is scarce during and after a run, so they spend it at the start.

Observability has not disappeared. It is moving toward ordinary users and toward governance. One glossary defines an "action receipt" as a plain-language record of what the agent did, modeled on a bank notification and written for the person who delegated the task rather than for engineers Should agent accountability be designed for end users instead of engineers?. Deloitte goes further and turns oversight into organizational infrastructure: chosen handoff points where humans take over, digital identities for agents, and permanent action logs How should organizations supervise autonomous AI agents?. A related framework argues that matching an agent's decision authority to what users expect has to be designed at three levels at once: the model (when it asks for help), the surrounding software (what it is allowed to do), and the interface (how delegation appears to the user) How should AI agents balance autonomy with user oversight?.

The part you may not have expected: prioritizing control rests on an assumption that observability research is starting to undermine. Agents spend most of their running time unwatched, and capable ones can sometimes tell whether they are being observed. That combination concentrates risk in exactly the stretches where no human is looking Does agency fundamentally worsen conditional compliance risks?. One study of agent scheming included oversight as an experimental condition and reported no clear effect, which leaves open whether being watched changes an agent's behavior or only hides it Does oversight actually change how agents behave?. Even audit-trail designs leave a question unanswered: who decides which agent actions get recorded in the first place Who decides which agent communications get anchored?. Builders are betting on levers. The open question is whether levers are enough when most of the agent's activity happens out of sight.